For time-resolved spectrum analysis of a voice signal, which is a method of voice information extraction used in voice recognition units and voice analyzers, the vocal frequency range is conventionally divided into channels by means of band-pass filters. The output power of each band-pass filter is rectified and smoothed with a specified time constant, and the output of all the band-pass filters therefore provides a correspondingly time-resolved segmented frequency analysis of the voice data. Such a device requires as many band-pass filters as the number of channels required for voice analysis or voice recognition. Thus, a dozen or more of these expensive band-pass filters may be required.